Thanks for this new feature. Here are 2 use cases I’d appreciate to see covered in the future:
Long document summary
Today the limitation in tokens is too strict for my usage. When I want a summary, it’s because the original document is quite long (eg : an R&D document 5 pages long).
Smart questions on all my workspace content
We store a lot of textual information in documents and rich text fields of entities about how our software is supposed to work.
I’d love if the AI assistant would be able to “index” this information (daily probably) and we would be able then to ask it questions and get answer based on our own data, rather than on the “standard indexes” of openai.
3 Likes
Yes would love to be able to do things like this:
1 Like
Thanks for implementing this - being able to connect databases is great - in my given example it struggles with referencing correct date ranges though - i ask for 2023 but it gives 2022 links. Maybe being able to choose which fields it can index could help?
It’s a limitation to how the RAG systems work. Basically it’s retrieving the top 5 records that are the most relevant to your query and then uses their contents to answer your question. So if you ask questions that are beyond what those 5 records can answer it’s very likely going to not answer your question and/or hallucinate.
Asking a question that spans many records will require at a minimum giving the AI the ability to write it’s own QraphQL queries but more likely will also require it to write it’s own code to process those queries. Something like what the ChatGPT code interpreter or advance analysis tool, or whatever it’s called now.
@Lee_Denny
This is our goal for near future. Indeed right now the chatbot knows nothing about Fibery structure, time, and other things.
2 Likes
